[WIP] Refresh MiniMax-M3 NVFP4 B200 8k1k disaggregated Dynamo-vLLM / 刷新 MiniMax-M3 NVFP4 B200 8k1k 分离式 Dynamo-vLLM 配置

[jasonlizhengjian]

Summary

• Carry forward the six non-speculative MiniMax-M3 NVFP4 B200 8k1k disaggregated Dynamo-vLLM recipes from [NV] feat(minimaxm3): add B200 disaggregated configs #2230 in a main-based PR so the benchmark sweep can run.
• Use vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9 instead of [NV] feat(minimaxm3): add B200 disaggregated configs #2230's vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2 image in the master configuration and all six recipes.
• Preserve [NV] feat(minimaxm3): add B200 disaggregated configs #2230's DEP2-prefill/TP4-decode topology, Dynamo 1.3.0.dev20260710, spread-worker placement, and concurrency sweep unchanged.
• Exclude EAGLE/MTP configurations.

Validation

• python -m pytest utils/matrix_logic/ utils/changelog_gate_tests/ -q: 299 passed.
• utils/validate_perf_changelog.py --base-ref origin/main --head-ref HEAD: passed; final newline present and matrix generated.
• The main-based changelog matrix contains exactly six 8k1k throughput jobs and six eval jobs, all with spec-decoding: none and the refreshed image.
• git diff --check origin/main...HEAD: passed.

中文说明

• 基于 [NV] feat(minimaxm3): add B200 disaggregated configs #2230 的六个非投机 MiniMax-M3 NVFP4 B200 8k1k 分离式 Dynamo-vLLM 配置创建面向 main 的 PR，以便触发基准测试扫描。
• 在主配置及全部六个配置中使用 vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9，替换 [NV] feat(minimaxm3): add B200 disaggregated configs #2230 中的 vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2 镜像。
• 保持 [NV] feat(minimaxm3): add B200 disaggregated configs #2230 的 DEP2 预填充/TP4 解码拓扑、Dynamo 1.3.0.dev20260710、分散式 worker 放置策略及并发扫描范围不变。
• 不包含 EAGLE/MTP 配置。

验证

• python -m pytest utils/matrix_logic/ utils/changelog_gate_tests/ -q：299 项通过。
• utils/validate_perf_changelog.py --base-ref origin/main --head-ref HEAD：通过；文件末尾换行及矩阵生成均正常。
• 基于 main 生成的变更日志矩阵恰好包含六个 8k1k 吞吐量任务和六个评估任务，全部使用 spec-decoding: none 及新镜像。
• git diff --check origin/main...HEAD：通过。

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[claude]

Claude finished @jasonlizhengjian's task in 1m 47s —— View job

---

Review of PR #2311

• Gather context (diff vs origin/feat/minimaxm3-b200-vllm-disagg)
• Verify master config + perf-changelog rules
• Verify image accessibility and recipe consistency
• Validation checks (see note below)
• Post review

LGTM - no blocking issues found

Verified against the stacked base feat/minimaxm3-b200-vllm-disagg:

• The diff is exactly the image bump — configs/nvidia-master.yaml (minimaxm3-fp4-b200-dynamo-vllm) plus all six non-speculative 8k1k recipe YAMLs move from vllm/vllm-openai:cu129-nightly-8e981630c... to vllm/vllm-openai:nightly-4080263bb...; topology, Dynamo version, and concurrency sweeps are untouched.
• perf-changelog.yaml was updated alongside the master config, the new entry is appended at the very end of the file, references the correct config key, and the file keeps its final newline.
• No stale references to the old image remain anywhere in the repo (the only occurrence is in the changelog text, which is expected).
• The new image is a public Docker Hub reference — no local paths or .sqsh files.

Note: this review sandbox blocks running pytest/network calls, so I could not independently rerun the test suite or confirm the Docker Hub tag exists — CI will exercise both, and the PR description reports 299 tests passing locally.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30022318630
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30022318630

[jasonlizhengjian]

Closing because MiniMax-M3 NVFP4 B200 disaggregated Dynamo-vLLM is not an existing recipe family on main, so this was an unintended new-family addition rather than a container refresh.